technique

Policy Gradient Methods

techniqueactivepolicy-gradient-methods-ef95a80a·2 events·first seen 28d ago

Aliases: Policy Gradient Methods

Co-occurring entities

OpenAI Entropy Regularization Soft Q-Learning Action-Dependent Factorized Baselines Variance Reduction

More like this (12)

policy gradient Evolved Policy Gradients Wasserstein Policy Gradient Proximal Policy Optimization Dual-Evidence Gradient Purification diffusion-based policy gradient accumulation Gradient Labs GRPO (Group Relative Policy Optimization)Integrated Gradients Gradient-Guided Reward Optimization Denoising Diffusion Policy Optimization

Recent events (2)

5Openai Blog·28d ago·source ↗

Equivalence between Policy Gradients and Soft Q-Learning

OpenAI published a research result establishing a formal equivalence between policy gradient methods and soft Q-learning, two major families of reinforcement learning algorithms. The work shows that under entropy regularization, these approaches are mathematically equivalent, unifying previously separate lines of RL research. This has implications for algorithm design, theoretical understanding, and the development of hybrid RL methods.

Alignment and RLHF Policy Gradient Methods Entropy Regularization OpenAI +1 more

3Openai Blog·28d ago·source ↗

Variance Reduction for Policy Gradient with Action-Dependent Factorized Baselines

OpenAI published a research paper on variance reduction techniques for policy gradient methods in reinforcement learning. The work introduces action-dependent factorized baselines as a way to reduce variance in policy gradient estimates without introducing bias. This is a foundational RL training methodology contribution relevant to improving sample efficiency in reinforcement learning.

Alignment and RLHF Action-Dependent Factorized Baselines Policy Gradient Methods Variance Reduction +1 more